Maybe the solution is to cull the bad, cliché writing from the training data.